Classification And Detection Of Mobile Phone Glass Defects Based On Deep Learning

With the proposal of "Made in China 2025" action program,China is rapidly moving towards the intelligent era of industrial 4.0.As a portable wireless communication tool,smart phone is developing towards entertainment,intelligence and convenience.The quality inspection and monitoring of mobile glass products is one of the important links in the production process of smart phones,especially the Cover Glass and TFT-LCD included in the touch screen of mobile phones.Because of the complexity of the defect of the protective screen,it is difficult to extract the defect features of the protective screen by the traditional machine visionbased detection algorithm.At present,most of the major smartphone manufacturers use a large number of people to detect the surface defects of the protective screen with naked eyes,but do not adopt the quality control detection algorithm based on machine vision.Long-term naked eye detection process is easy to cause visual fatigue,combined with subjective identification and training level,which makes the detection rate of artificial naked eye detection about 80%,which is difficult to meet the large-scale production needs of smart phones.At present,many scholars at home and abroad have proposed various TFT-LCD appearance defect detection algorithms for display screens.However,due to the development of mobile display screens in the direction of larger size,thinner thickness and higher resolution,traditional machine vision-based detection algorithms have higher requirements for optical systems when building detection platforms,such as illumination intensity,lighting angle,etc.In order to extract defects of high-resolution display screen,it depends on too many parameter settings,and can only detect the presence or absence of defects,so it is impossible to classify defects.In recent years,the feature extraction ability of deep learning for target detection,image segmentation,image recognition and other tasks has been proved to exceed the traditional machine learning algorithm in major international evaluations,and has become a research hotspot in the field of image detection,target recognition and large data analysis.Compared with the traditional machine vision algorithm,deep learning has a higher ability of feature extraction and classification for complex tasks,and has a broad application prospects in mobile glass surface defect detection.Therefore,this paper proposes a mobile phone glass surface defect classification detection algorithm based on in-depth learning.The main research contents and innovative work of this paper are as follows:The main research contents and innovative work of this paper are as follows:(1)Firstly,aiming at the problem of how to choose the suitable Convolution Neural Network(CNN)model for the surface defect image of mobile phone glass products,the advantages and disadvantages of different mainstream convolution neural networks are elaborated,and different networks are selected based on the sample of the surface image of mobile phone protection screen taken by high resolution linear array camera.The network model is validated by experiments.The features extracted from each layer of the visually designed CNN model are analyzed whether different network structures are sufficient to extract the non-linear features of the defects of the mobile phone protection screen.The appropriate network structure is selected to build a lightweight convolution neural network with smaller parameters while ensuring the detection effect.(2)Secondly,aiming at the features of TFT-LCD appearance defects of mobile phone display screen,a classification and detection algorithm of TFT-LCD defects of mobile phone display screen based on small sample learning is proposed.Through the analysis of TFT-LCD defect size morphology,the shallow features of input image are extracted with small-size convolution kernels of different sizes.Sparse convolution structure is introduced to enlarge the network width and depth,and then feature fusion is carried out to extract the deep features of image.Batch normalization technology is used instead of traditional normalization method to accelerate network convergence,and global average pooling is used instead.The deep features extracted by Full Connection Layer Fusion are sent to the Cross Entropy Loss Function for classification.(3)Thirdly,aiming at the incomplete extraction of non-linear feature of TFT-LCD surface defect by designed lightweight CNN in small sample,a model training method based on transfer learning is proposed.By loading the pre-training model of mobile phone protection screen defect detection,the initial parameters of the network are updated,the convergence speed of the designed network is accelerated,the occurrence of over-fitting is reduced,and the generality of the network is improved.It can reduce the recognition error rate due to improper selection of initial network parameters.Aiming at the problem that some kinds of defect images are less and the distribution of data is unbalanced,which affects the recognition accuracy,a generating sample training method based on deep convolution generation antagonism network(DCGAN)is proposed to ensure the balance between the size of training data and the distribution of data sets.Better test results can be obtained by sending the generated data into the migrated learning network model for intensive training.Then,aiming at the actual demand of industrial production line,a continuous training algorithm is designed.By locking the shallow feature extraction layer of the network,new samples are sent to the deep feature extraction layer to continue learning,and the training classifier is strengthened again.This method can avoid re-training caused by adding new training samples in industrial applications,and improve the practical application of the design algorithm.